[BUG] Split by regular expressions with ? and * repetition are not consistent with Spark

[NVnavkumar]

Describe the bug
We currently fall back to CPU for repetition quantifiers ? and * with split because the behavior is not consistent with Spark.

Steps/Code to reproduce bug
Example:
For the input string 31313 and the pattern 4?, split will produce ['3','1','3','1','3'] on CPU, and ['','3','1','3','1','3'] on the GPU.

Expected behavior
The behavior should be consistent with Spark so we can enable this on GPU.

Also, see #4468 for related issue regarding regexp_replace

[NVnavkumar]

Also, note the same behavior occurs with {0,} and {0,n} repetitions as well.

[NVnavkumar]

So relevant to solving this is to understand the JDK behavior. Here is the relevant line from doc for Pattern.split(...):

 When there is a positive-width match at the beginning of the input sequence then an empty leading substring is included at the beginning of the resulting array. A zero-width match at the beginning however never produces such empty leading substring.

Currently cuDF emulates Python behavior, which will always include an empty leading substring with a zero-width match.

[NVnavkumar]

Will file another issue specific to zero-width matching to capture this JDK edge case so that we can enable in other circumstances

So relevant to solving this is to understand the JDK behavior. Here is the relevant line from doc for Pattern.split(...):

 When there is a positive-width match at the beginning of the input sequence then an empty leading substring is included at the beginning of the resulting array. A zero-width match at the beginning however never produces such empty leading substring.

Currently cuDF emulates Python behavior, which will always include an empty leading substring with a zero-width match.

[NVnavkumar]

Filed #6958 to handle the zero-width match edge case. In other circumstances, we should enable * and ? on the GPU.